Support optional cupy for faster BSS evaluation

[sevagh]

Speed up BSS metrics with the GPU and CUDA-Python (cupy).

Regression tests pass. Usability is possibly a concern, given that clear_cupy_cache() should be run regularly between tracks to avoid GPU OOM errors when the FFT plan cache fills up for a single evaluated track.

I can squash my git commits as required.

[sevagh]

In hindsight I'm kinda uncertain where the "disable_cupy", "clear_cupy_cache", etc. functions would go, or if they even belong in the package.